An AI-assisted advice platform that treats personal data as sacred
Lead architect on an eight-platform digital engagement and advice estate, operating under POPIA and FAIS, where the language models are deliberately fenced in.
The problem
A South African health insurance and gap-cover provider wanted a digital engagement and advice estate spanning eight platforms. It operates under POPIA (data protection) and FAIS (regulated financial advice), so a wrong answer or a leaked record carries real consequences. AI was part of the brief, but it could not be trusted with the numbers.
What I did
- Wrote the programme’s technical architecture. After reassessing the client’s originally prescribed stack from first principles, I proposed a different direction, and the client formally adopted it as authoritative.
- Built the deployment platform from nothing: Ansible provisioning, single-node k3s per environment, a Helm umbrella chart, and Terraform for GCP
africa-south1. Staging, pre-production and production run on a build-once, promote-by-tag model. Schema migrations run as Helm pre-upgrade hooks and are rehearsed against restored staging data before merge. Seven services are live and wired for continuous delivery in staging. - Designed the Customer Interaction Store around one rule: every read is a write. The application role has no
SELECTon personal data. The only way in is a security-definer function that refuses to run unless an access-log row has been committed first. - Fenced the language models in: calculations are rules-only, and product figures come from a governed, citation-backed store.
- Designed a three-layer QA gate for the AI features. Separate agents handle authoring, review and test extraction. Pass-rate thresholds replace binary verdicts for probabilistic behaviour. Regression tests are keyed to a fingerprint of code, prompt, model, knowledge-store version and retrieval settings. A human signs off before anything ships.
How I showed up
The easy path was to build what had been specified. I thought the prescribed stack would cause problems later, so I made the case in writing, from first principles, and the client adopted it. That only works if the reasoning is visible and the people paying for it can follow it. I also stay in the code, so the architecture I recommend is one I’m prepared to build and run.
Outcome
Work is in progress. The architecture is the programme’s adopted direction, the three-environment platform is up, and seven services are continuously deployed in staging.